
For a while, AI video was a spectacle. A prompt produced a six-second clip of a shark swimming through a flooded office, everyone shared it, and nobody asked what happened next. That phase is over. Models such as Sora, Runway, Kling, Luma, Pika and Veo now generate footage that can survive a few seconds on a real screen: a commercial insert, a title sequence, a social hook, a pre-visualisation of a scene that will later be shot properly.
The interesting question is no longer "what can the model make?" It is "where does the generated clip sit inside a production that has a client, a deadline, a budget and a legal review?" That shift turns a creative toy into an operational problem, and operational problems are solved with workflows, not with better prompts alone.
A workflow here means a repeatable sequence: brief, visual bible, shot list, prompt architecture, generation passes, selection, assembly, sound, colour, delivery, archive. Generative models touch every stage and replace none of them. Teams that treat AI video as a magic render button end up with beautiful orphan clips and a frustrated editor. Teams that treat it as one new department inside an existing pipeline ship faster than they did before.
This guide lays out that pipeline in practical terms: how to plan shots that survive generation, how to hold a character together across scenes, how to choose between models, how to budget the unpredictable parts, and how to keep the legal side clean. It is written for producers, editors, motion designers and creative directors who want a system rather than a folder of prompts.
Traditional production spends most of its energy on logistics: locations, permits, crew, weather, talent, transport. AI-heavy production spends that energy on specification. A shot exists in your head, then in a document, then in a prompt, then in a latent space, then in a file. The cost of a wrong decision moves from the shoot day to the writing stage. If the description of a shot is vague, you will pay for that vagueness every time you press generate.
Every current model has failure modes you should plan around rather than discover at delivery. Hands interacting with small objects, text rendered inside a scene, complex multi-person choreography, precise camera moves that must match a live plate, and long uninterrupted takes all remain risky. Physics that looks right for two seconds can melt in the third. Reflections and shadows drift. Mirrors are a coin flip.
The practical response is shot design, not stubbornness. Break risky actions into shorter beats. Frame around hands. Avoid on-screen text — add it in post, where you can control the font. If a move must match existing footage, generate the move separately and composite.
When generating is cheap, choosing becomes expensive. A single scene can produce forty variants, and someone has to watch all of them. The teams that move fastest build a selection ritual: generate in batches by shot, review at thumbnail speed, promote only clips that pass a technical check for warping, flicker, limb integrity and continuity, then log why a clip was rejected. That rejection log becomes the most valuable document in the project, because it tells you which prompts to stop using.
Before generating anything, produce a one-page visual bible: palette, lens language, grain, aspect ratios, reference stills, motion attitude (handheld versus locked-off), and a written statement of what the piece must never look like. This document is the contract between client expectation and model output. It also prevents the classic failure of AI projects, where every shot looks like it came from a different film.
Build a shot list with columns for duration, camera, subject, action, lighting, audio intent and generation model. Write prompts into the same table so the shot list doubles as a production log. Include a continuity column: wardrobe, props, hair, time of day, screen direction. This column is what saves you when you return to a project after a week away.
Work in passes. Pass one explores composition and motion at low resolution or short duration. Pass two locks performance and camera. Pass three produces final candidates at delivery resolution. Between passes, update your prompt template with whatever worked. Treat prompts as versioned assets: hero-shot-v1, hero-shot-v2, with a note about what changed.
Cut AI footage exactly as you would cut live footage. Add sound design early, because sound is what makes generated motion feel intentional. Use stabilisation, grain matching and subtle camera shake to unify clips from different models. Colour grade last, with a shared look applied across the whole timeline so model-specific colour signatures disappear.
A reliable prompt has five parts, in this order:
Example: Medium tracking shot, 35mm, slow push-in. A woman in her forties in a charcoal wool coat, short dark hair, weathered hands. She walks toward a window and stops, one hand rising to the glass. Dim interior, north-facing light, dust in the air, overcast city outside. Muted teal and grey palette, fine grain, shallow depth of field, 16:9.
One action per prompt beat is the rule that most improves output quality. Models handle a single continuous movement far better than a sequence of instructions.
Keep a reusable continuity block, pasted into every prompt for a given scene: character identifiers, wardrobe, props, colour temperature, screen direction. Keep a second block of negative constraints: no on-screen text, no lens flare, no fast cutting, no extra people, no warped reflections. Negative constraints do not guarantee compliance, but they measurably reduce the number of unusable takes.
Character drift is the single most common complaint about AI video, and it is a workflow problem as much as a model problem. Four techniques reduce it:
Style consistency follows the same logic. Lock a look — palette, contrast curve, grain, focal length range — and apply it to every prompt. Never let a single shot carry a wildly different aesthetic without a narrative reason.
Different models are better at different things, and that difference is stable enough to plan around. Use this as a starting framework:
Evaluate candidates on your own footage rather than on marketing reels. Generate the same three test shots in every model: a medium shot of a person speaking, a fast lateral camera move, and a wide landscape with moving elements. Judge stability, texture, and how much post-processing each clip needs.
Mixing models is normal and often optimal, but it creates a continuity tax. Pay it deliberately: keep a single colour pipeline, standardise frame rates and aspect ratios on ingest, and place model-specific shots where their look difference reads as intentional (for example, memory sequences, inserts, or stylised cutaways). Never mix models randomly across a continuous scene.
AI production changes where money goes rather than eliminating cost. The traditional line items — crew, kit, location, catering, insurance — shrink or disappear. New line items appear: generation spend, iteration time, storage and transfer, upscaling, sound design to legitimise synthetic motion, and legal review.
Build budgets around iteration ranges rather than single numbers. Estimate the number of finished shots, then apply a ratio for attempts per shot based on how risky each shot is. Simple establishing shots might need a handful of attempts. Complex scenes with two interacting characters can need dozens. Multiply, add editing and finishing days, and you have a defensible quote.
The hidden cost is human attention. Reviewing hundreds of clips takes real hours. If you have a large volume of exploration, assign a dedicated reviewer and a strict triage rule: a clip is either promoted or discarded within seconds, and anything promoted gets a timestamped note. This keeps selection from eating your margin.
Three areas deserve attention before a project begins.
Likeness and voice. Do not generate a recognisable person without written permission. This includes near-lookalikes and stylistic imitations of famous performers. Keep signed releases for any real person referenced, and document the provenance of any facial reference images used.
Training data and rights. Understand what the platform's terms say about commercial use, ownership of outputs, and indemnification. Ask your client's legal team early, especially for broadcast, paid media and regulated industries.
Truthfulness. If a clip could be mistaken for a record of real events — news, documentary, testimony, product behaviour — label it clearly and keep a written record of how it was made. In advertising, distinguish between a stylised depiction and a claim. The safest habit is a short internal AI disclosure sheet per deliverable: which shots were generated, with which tool, and what was done in post.
Generating before the shot list exists. You will produce beautiful clips that do not cut together. Write the list first, always.
Rewriting prompts from scratch after every failure. Change one variable at a time and log it. Otherwise you learn nothing.
Ignoring audio until the end. Sound design is the fastest way to make generated footage look deliberate. Start it during assembly, not after picture lock.
Chasing perfect single takes. A slightly imperfect take that cuts well beats a flawless clip that does not fit the rhythm of the scene.
Skipping upscaling and finishing. Native model output rarely matches delivery standards. Plan for upscale, denoise, grain and grade, and budget render time for it.
Over-automating continuity. No tool currently guarantees a character across a long sequence. Plan coverage the way you would on a shoot: multiple angles, inserts, cutaways, and a backup framing that hides drift.
Can AI video replace a live shoot entirely?
It can carry certain formats — social spots, explainers, stylised sequences, pre-visualisation — on its own. Projects requiring precise performance, choreography, product accuracy or legal documentation still benefit from a hybrid approach in which generated shots support live footage rather than replace it.
How many attempts should I plan per finished shot?
For low-risk shots, expect single digits. For complex shots with characters, hands or matching moves, plan for many more and quote accordingly. Track actual ratios on your first projects so future estimates get tighter.
Do I need a different workflow for vertical social formats?
The pipeline is the same, but shot design changes. Vertical video rewards tighter framing, faster motion and strong opening frames. Generate natively in the target aspect ratio rather than cropping, because cropping often destroys intended framing.
What is the best way to keep a character consistent?
Combine a reference sheet, a verbatim descriptor block, consistent lighting language and a shared grade. Accept small drift and hide it with framing choices instead of re-generating indefinitely.
How should I present generated work to a client?
Lead with the finished cut, not the process. Then be transparent about which shots were generated and how, and provide the disclosure notes your client's legal team needs. Confidence about the workflow prevents the conversation from becoming a debate about authenticity.
Is it safer to use one model for everything?
One model simplifies continuity but limits quality on specific shot types. Most mature pipelines use a primary model plus one or two specialists, with a standardised finishing process that unifies the result.
What should be archived at the end of a project?
Prompts with versions, reference stills, selected and rejected clips, model and settings per shot, and the disclosure sheet. This archive is what makes the next project faster and cheaper, and it is the difference between a one-off experiment and a production capability.

Alexander
A practical guide to AI-assisted video production, from brief and shot list to prompt design, audio, editing, and final quality control.
Build a reliable free video transcription and editing workflow with practical tool choices, caption tips, quality checks, and AI handoff points.
Learn a reliable automated video transcription workflow with AI, from audio prep and speaker labels to quality checks, captions, and searchable transcripts.